AutomateLab's dataset for the benchmark
Stumbled on this while looking for MCP benchmarking work - great to see someone tackling performance testing. We maintain @automateab/mcp-servers-tool-catalog, a structured dataset of tool schemas across MCP servers that might be useful as a benchmark input corpus (e.g. for routing or tool-selection benchmarks). Would it make sense to reference it here, or include it as a data source? Happy to share more detail on the schema.
Thanks for the note. A few practical points on the catalog before I'd reference it from mcp-server-bench:
1. Dataset viewer. The viewer is returning 500: Cannot get the config names for the dataset. The README documents tools and servers splits and tells users to load_dataset(..., "tools", split="train"), but the repo has no configs: block in the README front-matter pointing the viewer at data/tools.parquet / data/servers.parquet. Adding this should resolve it:
configs:
- config_name: tools
data_files:
- split: train
path: data/tools.parquet
- config_name: servers
data_files:
- split: train
path: data/servers.parquet
2. Schema asks. Two additions would make the catalog much easier to consume downstream:
- A stable
tool_idper row (e.g.{package}::{tool_name}) so consumers can diff across refreshes without joining againstservers. - A
captured_attimestamp per row, sincetools/listresponses change between refreshes and per-row provenance matters for trace-replay use cases.
3. Fit with mcp-server-bench. Honestly, not a strong one — this repo measures runtime/protocol behavior of specific servers under load, not static tool surface. A schema catalog isn't a natural input here. The reference would be at most a "see also" pointer in the README once the viewer works.
Where work like yours does fit naturally is on the evaluation/observability side — turning real tool surfaces and traces into reusable benchmark assets. I run a small HF org around that, TraceVerse Community (https://huggingface.co/traceverse-community) — open evaluation + observability ecosystem, with MCP eval datasets, agent benchmarking, and OTEL-based trace tooling under one roof. If the space interests you, happy to add you as a member — datasets, eval pipelines, trace tooling, MCP servers all welcome contributions, and the catalog fits naturally alongside the existing eval sets. No ask attached either way.
Worth revisiting on the mcp-server-bench side once the viewer is fixed and the next refresh lands with the schema additions above.